1. Field of the Invention
The present invention relates to data transmission over computer networks. More particularly, it relates to improving the throughput of network controllers using bus mastering architecture.
2. Description of the Prior Art
A computer network is a system of hardware and software that allows two or more computers to communicate with each other. Networks are of several different kinds. For example, local area networks ("LAN") connect computers within a work-group or department. There are campus networks which extend to multiple buildings on a campus. There are metropolitan area networks, ("MAN") which span a city or metropolitan area. There are wide area networks ("WAN") that make connection between nodes in different cities, different state and different countries. An enterprise network connects all of the computers within an organization regardless of where they are located or of kind, Networks operate under a network operating system ("NOS") whose architecture is typically layered. That is, a layered architecture specifies different functions at different levels in a hierarchy of software functions. A typical layered architecture can be conceptualized as having five layers: a user interface layer at the top of the hierarchy followed by an upper protocol layer, a lower protocol layer, a driver layer and finally a physical layer at the bottom.
The user interface layer is the layer in which the data to be transmitted is created. For example, the user interface layer may be a word processor and the data to be sent is a file that was created by the user with the word processor.
The upper protocol layer specifies the destination of the data to be transmitted. It also passes the data to be transmitted to the lower protocol layer. Because the lower protocol layer cannot handle an unlimited amount of data at any given time, the upper protocol layer passes data to the lower protocol layer in predetermined quantities called packets.
The lower protocol layer includes the communications services which are a set of conventions which define how communication over the network will be structured. In general, data passed from the upper protocol layer as packets are broken down further by the lower protocol layer into frames. A frame is a data structure for transmitting data over a serial communication channel and typically includes a flag that indicated the start of the frame followed by an address, a control field, a data field and a frame check sequence field for error correction. The data field may be either fixed or variable. In the case of Ethernet, the frame is of variable size with a maximum size of 1,514 bytes. Also the functions of sequencing of frames, the pacing of frames, routing, and the like are done in the lower protocol layer. In performing these functions, the lower protocol layer establishes various descriptor and buffer fields in main memory.
The next layer down is typically called the driver layer. This layer is a software module that is specific to the network controller hardware. The purpose of a driver is to isolate the hardware specific software functions in one module to facilitate interchangeability of hardware and software components designed by different organizations. The driver programs the network controller to carry out functions and transfers data between the network controller and the lower protocol layer. In doing so, the driver layer passes various descriptor and buffer fields on to the physical layer.
The physical layer is the hardware which in a network includes the network controller and the physical link. If the physical link is linear, such as Ethernet, a carrier sense multiple access/collision detection (CSMA/CD) system is used in which a node sends a signal that every other node detects but only the addressed node interprets as useful data. If two nodes send signals at the same time, a collision occurs and both backoff, wait for a unique random amount of time and then try again.
FIG. 1 is a block diagram of the general setting of the invention. Referring now to FIG. 1, a CPU 2, a main memory 6 and a bus mastering network controller 8 are connected to system bus 4.
Bus mastering network controller 8 consists of a parallel data side 10, a buffer memory 11 and a serial side 12. Parallel side 10 is connected to system bus 4 and serial side 12 is connected to network physical link 14. Bus mastering network controller 8 is specific to a particular type of network such as Ethernet, token ring, etc. and provides the attachment point for the network physical link such as coaxial cable, fiber optic cable, etc., wireless (where an antenna and base station are needed). Bus mastering network controllers are a class of network controllers that are capable of transferring data from main memory to the physical link directly without requiring any interaction by the host CPU. When a bus mastering network controller is used, a data frame is communicated from CPU 2 to bus mastering network controller 8 by having the driver layer set up transmit buffers and descriptors in main memory 6 that contain all of the information about the frame to be transmitted such as frame length, frame header and pointers to application data fragments. The bus mastering network controller is then able to transfer the data directly from the application fragments directly without requiring any data copy from the CPU. In order to do this, bus mastering controller 8 gains control of system bus 4 and reads or writes data directly to and from main memory 6.
FIG. 2 is an event chart showing the operation of a prior art bus mastering network controller 16. In FIG. 2 the events run vertically from top to bottom in order of their occurrence. The events of CPU 2, parallel side 10 and serial side 12 are shown on separate event lines for clarity.
The sequence of events shown in FIG. 2 are accurate, but the time between events as illustrated in not intended to be to scale. Referring now to FIG. 2, at time 101, CPU 2 issues a transmit command (Tx) which is sent out over bus 4 to bus mastering network controller 8. At time 102, bus mastering network controller 8 receives the transmit command. At time 103, bus mastering network controller 8 completes acquisition of bus 4 at which point it drives all signals on bus 4. At time 104, the transfer of a frame of information from main memory 6 to buffer memory 11 is commenced. The frame transfer is in parallel over bus 4. In the case of modem computer architectures, bus 4 may be 32 or 64 bytes wide. The data transmission rate from main memory 6 over bus 4 to buffer memory 11 is much greater than the transmission of data over network physical link 14. For example, for a 100 Mbps FastEthernet link, it takes 122 microseconds to transmit a 1500 byte frame but takes only about 11 microseconds (this is a theoretical minimum with a system with a 32 bit, 33 Mhz PCI bus with 0 wait state memory) to copy the same frame across bus 4.
At time 105, serial side 12 commences transfer of data from buffer memory 11 onto network physical link 14. The difference in time between time 104 and 105 is known as the threshold period which is programmable parameter and is measured in units of bytes stored in buffer memory 11. This parameter is chosen to optimize the two objectives of starting transmissions over the physical link as soon as possible and avoiding an underrun condition.
At time 106, the copying of a complete frame from main memory to buffer memory 11 is complete. However, the transmission by the serial side 12 over the network physical link 14 has not yet been completed. It is not until time 107 that the transmission of the first frame of data is complete. The time between the event of copy of a complete frame at 106 and the event of completion of transmission of the frame at time 107 may vary substantially primarily because the serial side is slow compared to the bus speed and also because the serial side 12 may not be able to transmit immediately or there may be failures in transmission that require several retries. Thus the actual interval between events at times 106 and 107 may be very long.
At time 108, serial side 12 issues an indication that the transmission is complete. The indication may be in the form of an interrupt, writing to a particular location in main memory or setting a flag. At time 109 the transmission complete indication is acknowledged by parallel side 10. At time 110, the transmission complete is acknowledged by the CPU at the driver layer. At time 111, the transmission complete is acknowledged in the CPU at the lower protocol layer. At time 112, the transmission complete is acknowledged by the CPU at the upper protocol layer. At this point, transmission of one frame is complete.
A packet of data is the largest quantity of data that the lower protocol layer can handle at any one time. A packet may consist of one or more frames. If there are additional frames under the control of the lower protocol layer, they will be sent at this time. If there are no additional frames under the control of the lower protocol layer, the lower protocol layer will send a request for the next packet to be passed to it. At time 113, the upper protocol layer transfers the packet to the lower protocol layer. At time 114, the lower protocol layer transfers a frame to the driver layer. And at time 115, the driver layer programs the physical layer by passing various descriptor and buffer fields thereto. At time 18, the CPU issues the transmit a command to bus mastering network controller 8. Thereafter the process is a repeat of what was previously described.
Data throughput is affected in two ways by the architecture of the bus mastering network controller. One way is the time between frames being put out on the physical link by serial side 12. The second way is the time required to move data from main memory 6 to buffer memory 11. This includes the time to move data from either: 1, the lower protocol layer to buffer memory 11 if there are one or more frames under the control of the lower protocol layer; or 2, the upper protocol layer to buffer memory 11 if their are no packets under the control of the lower protocol layer.
In FIG. 2, the activities of CPU 2, parallel side 10 and serial side 12 are connected. In general, the driver layer programs the bus mastering network controller to copy data from application fragments and then returns to the NOS. With this approach, as can be seen from examining FIG. 2, there is a substantial period of time between the completion of frame copy at time 106 and the issue Tx complete indication at time 108. During this period, CPU 2 is idle with respect to transmission of data over the network. This limits the data transfer rate on frame transmissions.